A deep learning-based radiomics model for noninvasive diagnosis of melanoma.

To develop a noninvasive diagnostic model integrating deep learning and radiomics for improving the accuracy and clinical utility of early melanoma diagnosis. A total of 350 patients with cutaneous pigmented lesions admitted to our hospital between January 2022 and December 2024 were retrospectively enrolled and randomly divided into a training set ( n  = 245) and a validation set ( n  = 105) in a 7:3 ratio. Complete information were obtained for all patients. Univariate analysis was used to screen factors associated with malignant melanoma. Variables were refined using the least absolute shrinkage and selection operator regression, and independent predictors were identified via multivariate Logistic regression. Random forest (RF), support vector machine (SVM), and K-nearest neighbors (KNN) models were constructed using Python 3.8.5 and the sklearn library. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). Results from univariate analysis and multivariate logistic analysis showed that lesion diameter, entropy (first-order statistic), long run emphasis, large area emphasis, wavelet contrast, wavelet energy, and the ResNet50-layer49 output were independent risk factors for malignant melanoma (all P  < 0.05). The AUC of the RF model (0.794) was significantly higher than that of the KNN algorithm model (0.755) and the SVM model (0.768), making it the optimal model. The RF model constructed based on deep learning-based radiomics features can be effectively applied to the noninvasive diagnosis of melanoma in patients with cutaneous pigmented lesions. Among these features, entropy (first-order statistic), long-run emphasis, and wavelet contrast are the key predictive indicators.